Research On Document-level Event Extraction Methods In Financial Domain

With the development of deep learning technology and the mining of massive data,event extraction has received widespread attention from domestic and foreign researchers.The task of document-level event extraction in the financial field aims to extract structured event descriptions from texts related to financial investment,intelligent decision-making,financial management,etc.,and provide important data support for financial regulation,public opinion guidance,and social governance.The challenge of this task lies in the complexity of the text and event structure.The complexity of the text is due to the scarcity of existing annotated data for document-level events,the difficulty of manual annotation,and the long-tailed distribution of different events,resulting in sparse document-level event data.The complexity of the event structure lies in the fact that a document contains multiple events,and event arguments are distributed in different sentences.The cross-sentence discreteness of arguments hinders the model from encoding long-range event semantic information,resulting in incomplete argument recognition.To address these two types of problems,the main research contents of this paper are as follows:To address the problem of sparse document-level event data,this paper proposes a document-level event extraction method based on frame semantic mapping and type awareness.First,the Chinese Frame Net(CFN)is mapped to Chinese document-level events,and the definition information of trigger words is integrated into the text input layer.A sliding window mechanism is used to fully perceive the context to alleviate the impact of limited text input.Secondly,in the argument recognition stage,the event type is used as the category label and concatenated with the document-level text as input features.This method transforms document-level multiple events into multiple single events to some extent,alleviating the phenomenon of overlapping arguments of multiple events.Additionally,adversarial training is introduced to enhance the model’s robustness.Experimental results on the Du EE-fin and CCKS2021 datasets show that this method has some improvement compared to current mainstream methods.To address the problem of scattered cross-sentence arguments,this paper proposes a document-level event extraction method based on heterogeneous graph neural networks and event subgraphs.First,sentence and entity information representations are obtained through entity recognition.Then,a document graph based on heterogeneous graph neural networks is constructed to capture the information between different entity nodes and edge attribute vectors.In addition,a graph attention network is introduced to divide events containing multiple related sentences,forming interactive event subgraphs in different events.This method captures the correlation between events and ensures the independence of each subgraph,further improving the impact of document-level multiple events.The effectiveness of this method is demonstrated on the CFA dataset.The main contributions of this paper include:(1)combining CFN with the task of document-level event extraction,exploring the relationship between frames and events,and formulating three types of constraints between frames and events.Two-level mappings of frames and events are constructed by introducing CFN as external knowledge.(2)proposing a document-level event extraction method based on frame semantic mapping and type awareness,improving the problem of sparse document-level event data and enhancing event extraction performance.(3)proposing a document-level event extraction method based on graph neural networks and event subgraphs,alleviating the problem of scattered cross-sentence arguments and improving the model’s cross-sentence argument integration ability.